EMG-Based Cross-Subject Silent Speech Recognition Using Conditional Domain Adversarial Network
Yakun Zhang, Huihui Cai, Jinghan Wu, Liang Xie, Minpeng Xu, Dong Ming, Yan Ye, Erwei Yin
Abstract
Machine learning techniques have achieved great success in electromyography (EMG) decoding, but EMG-based cross-subject silent speech recognition (SSR) received less attention because of its high individual variability. Therefore, this article explores the field of cross-subject SSR to improve the recognition performance of EMG data collected from new subjects. First, this article reports on applying time-series features and 1-D convolutional neural networks (1D-CNNs) for cross-subject SSR. Second, this article proposes using a conditional domain adversarial network (CDAN) to solve the problem of reduced cross-subject SSR accuracy in the few samples’ data sets. It innovatively integrates the maximum mean difference (MMD) loss to get an improved CDAN (ICDAN). While 1D-CNN is a feature extraction network that can meet the needs of cross-subject SSR in large data sets, the recognition effect will be weakened in small data sets. Adding an ICDAN network after the feature extraction network can improve the problem of data distribution differences between the two domains, and further enhance recognition performance. The results show that the 1D-CNN model based on time-series features yields better results in the SSR of new subjects, and the ICDAN model can further improve the classification accuracy of cross-subjects in a few sample data sets by 14.88%.